Set dunder operators too permissive?

Similar to #1840

With #3181 I made KeysView consistent with Set but I really think the operators are confusingly permissive.

Let's look at them

# difference
{'1'} - {2}
reveal_type({'1'} - {2})  # => Set[str]

This should fail because this is always a null operation and thus
a programmer error.

# intersection
{'1'} & {2}  # => Set[str]
reveal_type({'1'} & {2})  # => Set[str]

This really should fail because it is unlikely meaningful; the
return type is always an empty set thus it is also not a Set[str]

# sym difference
{'1'} ^ {2}
reveal_type({'1'} ^ {2})  # => Set[Union[str, int]]

Although the assumed union type is correct, Set[A] & Set[B] => Set[A|B] just doesn't make sense. The operator can only meaningful be applied to same type sets. What we do here is just a form of plain concat or union.

# union
{'1'} | {2} # MAYBE OK
reveal_type({'1'} | {2})  # => Set[Union[str, int]]

If we look at Sets as just containers without some
properties, we could accept it to construct mixed
type Sets using 'union'. I think it is okay for Python,
the name union makes it meaningful, but I maybe never did
that in practice.
Since it adds to the type it is also likely that
the following part of the program catches mistakes.

(But lists don't allow that [1] + ['2'] although they're really
arbitrary containers. Except my PR #3183.)

Start with this issue's examples and compare the related discussions in #1840, #3181, and #3183. Determine which Set and KeysView operators should reject mixed element types, including whether union remains allowed. The work is done when the project has an agreed rule for each operator and the relevant typing behavior reflects that decision.